Apparatus and Method for Generating a Dialogue State Tracking Model

1. An apparatus for generating a dialogue state tracking model, comprising: a storage, being configured to store a database; and a processor, being electrically connected to the storage and configured to retrieve a first field feature corresponding to a queried field from the database according to the queried field corresponding to a queried message, retrieves a first candidate-term feature for each of at least one first candidate-term corresponding to the queried field from the database, and integrates the at least one first candidate-term feature into a first integrated feature, wherein the processor further generates at least one relation sub-sentence of a reply message corresponding to the queried message, generates a sentence relation feature according to the at least one relation sub-sentence, generates a queried field related feature according to the first field feature, the first integrated feature, and the sentence relation feature, and trains the dialogue state tracking model according to the queried field related feature.

2. The apparatus of claim 1 , wherein the processor further integrates the first field feature and the first integrated feature into a first output feature, inputs the first output feature into a first neural network model to generate a second output feature, integrates the second output feature and the sentence relation feature into a third output feature, and inputs the third output feature into a second neural network model to generate the queried field related feature.

3. The apparatus of claim 1 , wherein the processor further retrieves a first reply feature corresponding to a field reply from the database according to the field reply corresponding to the queried message, retrieves a second candidate-term feature for each of at least one second candidate-term corresponding to the field reply from the database, and integrates the at least one second candidate-term feature into a second integrated feature, wherein the queried field related feature is generated by the processor according to the first field feature, the first reply feature, the first integrated feature, the second integrated feature, and the sentence relation feature.

4. The apparatus of claim 3 , wherein the processor further retrieves a third candidate-term feature for each of at least one third candidate-term from the database according to the reply message and a plurality of default fields, integrates the at least one third candidate-term feature into a third integrated feature, and generates an advanced relation feature according to the sentence relation feature and the third integrated feature, wherein the processor further analyzes the reply message by an NLU algorithm to generate a semantic field and a semantic reply, retrieves a second field feature corresponding to the semantic field from the database, retrieves a fourth candidate-term feature for each of at least one fourth candidate-term corresponding to the semantic field from the database, and integrates the at least one fourth candidate-term feature into a fourth integrated feature, wherein the processor further retrieves a second reply feature corresponding to the semantic reply from the database, retrieves a fifth candidate-term feature for each of at least one fifth candidate-term corresponding to the semantic reply from the database, integrates the at least one fifth candidate-term feature into a fifth integrated feature, and generates a semantic feature according to the second field feature, the fourth integrated feature, the second reply feature, and the fifth integrated feature, wherein the dialogue state tracking model is trained by the processor according to the queried field related feature, the advanced relation feature, and the semantic feature.

5. The apparatus of claim 3 , wherein the processor further integrates the first field feature and the first integrated feature into a first output feature, integrates the first reply feature and the second integrated feature into a second output feature, and inputs the first output feature, the second output feature, and the sentence relation feature into a neural network model to generate the queried field related feature.

6. The apparatus of claim 1 , wherein the processor further retrieves a second candidate-term feature for each of at least one second candidate-term from the database according to the reply message and a plurality of default fields, integrates the at least one second candidate-term feature into a second integrated feature, and generates an advanced relation feature according to the sentence relation feature and the second integrated feature, wherein the dialogue state tracking model is trained by the processor according to the queried field related feature and the advanced relation feature.

7. The apparatus of claim 6 , wherein the processor further integrates the sentence relation feature and the second integrated feature into an output feature and inputs the output feature into a neural network model to generate the advanced relation feature.

8. The apparatus of claim 1 , wherein the processor further analyzes the reply message by a Natural Language Understanding (NLU) algorithm to generate a semantic field and a semantic reply, retrieves a second field feature corresponding to the semantic field from the database, retrieves a second candidate-term feature for each of at least one second candidate-term corresponding to the semantic field from the database, and integrates the at least one second candidate-term feature into a second integrated feature, wherein the processor further retrieves a reply feature corresponding to the semantic reply from the database, retrieves a third candidate-term feature for each of at least one third candidate-term corresponding to the semantic reply from the database, and integrates the at least one third candidate-term feature into a third integrated feature, wherein the processor further generates a semantic feature according to the second field feature, the second integrated feature, the reply feature, and the third integrated feature, wherein the dialogue state tracking model is trained by the processor according to the queried field related feature and the semantic feature.

9. The apparatus of claim 8 , wherein the processor further integrates the second field feature and the second integrated feature into a first output feature, integrates the reply feature and the third integrated feature into a second output feature, integrates the first output feature and the second output feature into a third output feature, and inputs the third output feature into a neural network model to generate the semantic feature.

10. The apparatus of claim 1 , wherein the processor further retrieves a second candidate-term feature for each of at least one second candidate-term from the database according to the reply message and a plurality of default fields, integrates the at least one second candidate-term feature into a second integrated feature, and generates an advanced relation feature according to the sentence relation feature and the second integrated feature, wherein the processor further analyzes the reply message by an NLU algorithm to generate a semantic field and a semantic reply, retrieves a second field feature corresponding to the semantic field from the database, retrieves a third candidate-term feature for each of at least one third candidate-term corresponding to the semantic field from the database, and integrates the at least one third candidate-term feature into a third integrated feature, wherein the processor further retrieves a reply feature corresponding to the semantic reply from the database, retrieves a fourth candidate-term feature for each of at least one fourth candidate-term corresponding to the semantic reply from the database, integrates the at least one fourth candidate-term feature into a fourth integrated feature, and generates a semantic feature according to the second field feature, the third integrated feature, the reply feature, and the fourth integrated feature, wherein the dialogue state tracking model is trained by the processor according to the queried field related feature, the advanced relation feature, and the semantic feature.

11. A method for generating a dialogue state tracking model, being executed by an electronic apparatus, the electronic apparatus storing a database, the method comprising: (a) retrieving a first field feature corresponding to a queried field from the database according to the queried field corresponding to a queried message; (b) retrieving a first candidate-term feature for each of at least one first candidate-term corresponding to the queried field from the database; (c) integrating the at least one first candidate-term feature into a first integrated feature; (d) generating at least one relation sub-sentence of a reply message corresponding to the queried message; (e) generating a sentence relation feature according to the at least one relation sub-sentence; (f) generating a queried field related feature according to the first field feature, the first integrated feature, and the sentence relation feature; and (g) training the dialogue state tracking model according to the queried field related feature.

12. The method of claim 11 , wherein the step (f) comprises: integrating the first field feature and the first integrated feature into a first output feature; inputting the first output feature into a first neural network model to generate a second output feature; integrating the second output feature and the sentence relation feature into a third output feature; and inputting the third output feature into a second neural network model to generate the queried field related feature.

13. The method of claim 11 , further comprising: retrieving a first reply feature corresponding to a field reply from the database according to the field reply corresponding to the queried message; retrieving a second candidate-term feature for each of at least one second candidate-term corresponding to the field reply from the database; and integrating the at least one second candidate-term feature into a second integrated feature, wherein the step (f) generates the queried field related feature according to the first field feature, the first reply feature, the first integrated feature, the second integrated feature, and the sentence relation feature.

14. The method of claim 13 , further comprising: retrieving a third candidate-term feature for each of at least one third candidate-term from the database according to the reply message and a plurality of default fields; integrating the at least one third candidate-term feature into a third integrated feature; generating an advanced relation feature according to the sentence relation feature and the third integrated feature; analyzing the reply message by an NLU algorithm to generate a semantic field and a semantic reply; retrieving a second field feature corresponding to the semantic field from the database; retrieving a fourth candidate-term feature for each of at least one fourth candidate-term corresponding to the semantic field from the database; integrating the at least one fourth candidate-term feature into a fourth integrated feature; retrieving a second reply feature corresponding to the semantic reply from the database; retrieving a fifth candidate-term feature for each of at least one fifth candidate-term corresponding to the semantic reply from the database; integrating the at least one fifth candidate-term feature into a fifth integrated feature; and generating a semantic feature according to the second field feature, the fourth integrated feature, the second reply feature, and the fifth integrated feature, wherein the step (g) trains the dialogue state tracking model according to the queried field related feature, the advanced relation feature, and the semantic feature.

15. The method of claim 13 , wherein the step (f) comprises: integrating the first field feature and the first integrated feature into a first output feature; integrating the first reply feature and the second integrated feature into a second output feature; and inputting the first output feature, the second output feature, and the sentence relation feature into a neural network model to generate the queried field related feature.

16. The method of claim 11 , further comprising: retrieving a second candidate-term feature for each of at least one second candidate-term from the database according to the reply message and a plurality of default fields; integrating the at least one second candidate-term feature into a second integrated feature; and generating an advanced relation feature according to the sentence relation feature and second integrated feature, wherein the step (g) trains the dialogue state tracking model according to the queried field related feature and the advanced relation feature.

17. The method of claim 16 , wherein the step of generating the advanced relation feature comprises: integrating the sentence relation feature and the second integrated feature into an output feature; and inputting the output feature into a neural network model to generate the advanced relation feature.

18. The method of claim 11 , further comprising: analyzing the reply message by an NLU algorithm to generate a semantic field and a semantic reply; retrieving a second field feature corresponding to the semantic field from the database; retrieving a second candidate-term feature for each of at least one second candidate-term corresponding to the semantic field from the database; integrating the at least one second candidate-term feature into a second integrated feature; retrieving a reply feature corresponding to the semantic reply from the database; retrieving a third candidate-term feature for each of at least one third candidate-term corresponding to the semantic reply from the database; integrating the at least one third candidate-term feature into a third integrated feature; and generating a semantic feature according to the second field feature, the second integrated feature, the reply feature, and the third integrated feature, wherein the step (g) trains the dialogue state tracking model according to the queried field related feature and the semantic feature.

19. The method of claim 18 , wherein the step of generating the semantic feature comprises: integrating the second field feature and the second integrated feature into a first output feature; integrating the reply feature and the third integrated feature into a second output feature; integrating the first output feature and the second output feature into a third output feature; and inputting the third output feature into a neural network model to generate the semantic feature.

20. The method of claim 11 , further comprising: retrieving a second candidate-term feature for each of at least one second candidate-term from the database according to the reply message and a plurality of default fields; integrating the at least one second candidate-term feature into a second integrated feature; generating an advanced relation feature according to the sentence relation feature and the second integrated feature; analyzing the reply message by an NLU algorithm to generate a semantic field and a semantic reply; retrieving a second field feature corresponding to the semantic field from the database; retrieving a third candidate-term feature for each of at least one third candidate-term corresponding to the semantic field from the database; integrating the at least one third candidate-term feature into a third integrated feature; retrieving a reply feature corresponding to the semantic reply from the database; retrieving a fourth candidate-term feature for each of at least one fourth candidate-term corresponding to the semantic reply from the database; integrating the at least one fourth candidate-term feature into a fourth integrated feature; and generating a semantic feature according to the second field feature, the third integrated feature, the reply feature, and the fourth integrated feature, wherein the step (g) trains the dialogue state tracking model according to the queried field related feature, the advanced relation feature, and the semantic feature.